hypothesis
active
hypothesis:shaping-what-a-model-is-disposed-to-say-in-counterfactual-continuations-will-shape-what-it-silently-thinks-and-does-in-the-original-contextShaping what a model is disposed to say in counterfactual continuations will shape what it silently thinks and does in the original context.
Motivates reflection training.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Ablating 176 ethics lens vectors raises reflection-trained dishonesty from 0.07 to 0.22 leaving base unchanged; deception 0.05 to 0.23.associated_withCausal mediation.
- Behavioral gain.
- Implanted workspace.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Acknowledged limitation.
- All models exhibit above-baseline representation of the think word when instructed to think about itfinding0.772In the intentional control experiment, all tested models show above-zero cosine similarity to the think word's concept vector.
- Addresses skeptical alternative that reports reflect only conversational content
- Fig. 9c, Fig. 37.
- Key jailbreak insight.
- Unpredictability is a necessary condition for genuine adaptation.